Skip to content

v0.1.6

Choose a tag to compare

@github-actions github-actions released this 19 Sep 17:31
· 95 commits to main since this release
v0.1.6
fd63db4

Compatibility: verification is stricter

Some documents accepted by 0.1.5 are intentionally rejected by 0.1.6. Verification now rejects overflowing economic values, malformed JSON, nonnumeric trust summaries, wrong-profile documents, noncanonical raw-literal hashes and incorrectly rounded totals. Producers must correct and reseal invalid documents; the verifier does not repair them.

Equivalent float spellings normalize consistently across Python and JavaScript, and correctly rounded large finite receipts are accepted. Existing vector bytes and verdicts are preserved, with no hash-version change.

The stranger CLI reports VERDICT: PARSE_ERROR for malformed or overflowing JSON and retains exit code 1. Semantic failures and unreadable files still report INVALID. Consumers matching output text must recognize all three outcomes.

Read the compatibility decision before upgrading.

Changes

  • Close the token-detail privacy bypass before public bundle inputs are copied.
  • Preserve top-level billing tiers in savings reports; exclude conflicting or nonstandard tiers. Short-context pricing is a stated basis, not a lower bound on net cache savings.
  • Add OpenAI prompt-cache reporting, refresh output-price entries, exclude batch traffic, and support LiteLLM routing-savings receipts. Provider prompt-cache estimates remain report-only and self-reported.
  • Add a hardened, pinned LiteLLM integration CI job and receipt-level regressions, including genuine rounding ties, large values and Windows console encodings.
  • Expand the conformance corpus to 65 vectors and 75 checks, adding 27 vectors without modifying pre-existing vector documents. Require both verifiers to accept every fuzz base before mutation.
  • Correct unsupported agreement claims and update the spec and live verifier documentation.

Evidence and limits

The tagged release workflow verifies the signing key and runs the full release gate, including ten reproducibility rounds, before publication. Release assets include the gate report, SHA-256 checksums and build provenance. The wheel and source distribution uploaded to PyPI are checked against those same release checksums.

Passing the tests and sampled differential fuzz is bounded evidence, not proof of agreement on every input. The stdlib-only core and the integrity/magnitude/veracity distinction remain unchanged; no new claim of externally verified real-world savings is made.

Full changelog